On October 26, 2026, Anthropic filed its Form S-1, and for four days the public debate was dominated by valuation, revenue growth, and an $80.7 billion headline loss. Then the Financial Times reported something with more direct consequences for how the company will pay for its future: Blackstone, the anchor investor in a $63 billion financing deal for Anthropic's chips, has told investors in that deal that it expects the AI lab to receive a high junk rating after its IPO. This article walks through what that forecast means, why it matters, and what developers and infrastructure buyers should watch in the coming quarters.
The forecast that reversed
Blackstone had previously told investors it expected Anthropic to secure an investment-grade rating before 2027. According to the FT, that view has changed. The firm now expects a rating in the double-B range, which is the top tier of speculative grade. Moody's, S&P, and Fitch have not made an official determination. The forecast is a prediction, not a decision. But predictions from the institution underwriting the debt tend to travel, and this one tells you something about how the people closest to the deal are pricing Anthropic's risk.
Why did Blackstone change its view? Bankers working with both Anthropic and OpenAI have spent months trying to persuade the major agencies to hand the two AI developers investment-grade ratings once they are public. Blackstone itself had been telling investment partners that Anthropic would get there before the end of the year. The revision came after preliminary guidance from the agencies, pointing toward BB or higher within speculative grade.
That shift did not happen in isolation. Anthropic has been signing compute deals at a pace that makes the credit math more difficult. Its prospectus describes a plan to spend more than $800 billion on computing infrastructure through 2045. A company that commits to hundreds of billions in future payments is not going to be priced like a mature software business, regardless of how fast revenue is rising. The agencies are saying that out loud, at least in preliminary conversations.
The sequence matters more than the headline. Bankers have spent months trying to convince the agencies that Anthropic's revenue trajectory, its enterprise retention, and its cash pile should outweigh the still-negative accounting picture. Blackstone's revised guidance shows that those arguments are not yet landing at the rating-committee level. The agencies are not questioning whether the AI buildout is real. They are questioning whether a business with this much committed future spending has yet earned the right to be treated as a low-risk borrower.
What a double-B rating actually means
Credit ratings split the corporate world in two. Ratings from AAA down to BBB minus are investment grade. Ratings from BB plus down to D are speculative grade, often called high yield or junk. The boundary is not cosmetic. It decides which institutions are allowed to buy the bonds, how much collateral is required, and how much interest the issuer has to pay.
| Band | Category | Typical issuer profile |
|---|---|---|
| AAA to A | Investment grade | Large, profitable, low leverage, long track record |
| BBB plus to BBB minus | Investment grade | Stable cash flow, moderate leverage, some cyclicality |
| BB plus to BB minus | Speculative grade, top tier | Growing, leveraged, limited through-cycle history |
| B plus to B minus | Speculative grade | High leverage, thin margins, elevated default risk |
| CCC and below | Speculative grade | Distressed or near default |
An issuer with a double-B rating can borrow, but it borrows differently. The pool of buyers narrows to high-yield funds, hedge funds, private credit vehicles, and a sliver of crossover investors. Many insurance companies and pension funds operate under mandates that restrict or forbid meaningful exposure to speculative-grade debt. That is exactly the constraint the FT flagged: a missing investment-grade rating narrows Anthropic's lender base and raises its cost of capital.
Second-order effects matter as much. Ratings are public signals. A BB rating on Anthropic debt becomes a benchmark that counterparties cite in lease negotiations, in supplier contracts, and in the internal risk models of every company that has extended the lab credit backstops or capacity commitments. If you are running an engineering team that relies on Anthropic APIs, you might not look at the rating directly. The cloud providers and compute brokers that negotiate with Anthropic will. Their financing costs feed into what they charge for capacity, and those costs ultimately shape the price and availability of the models you use.
There is also a status effect. In the bond market, the difference between BBB minus and BB plus is the difference between a borrower that is eligible for conservative portfolios and one that is not. A company can be operationally strong and still fail that test. The rating agencies are slow to award investment grade to a company with no through-cycle record, even when its quarterly growth numbers look spectacular.
Inside the $63 billion deal that made the rating a public question
The rating question became urgent because of how the chip financing was built. On October 26, 2026, Bank of America, Blackstone, Citi, and Morgan Stanley announced a $63 billion financing structure to buy Broadcom AI accelerators and networking equipment for lease to Anthropic, executed through a newly formed special purpose vehicle. Deliveries are expected to start in 2027 under a previously announced compute capacity agreement involving Broadcom. The FT later described the transaction as Project North Star, covering Anthropic's lease of 1.5 gigawatts of Google-designed chip capacity.
The stack is layered, and each layer carries a different rating.
| Layer | Size | Backing | Rating signal |
|---|---|---|---|
| Senior (first lien) debt | About $43.5 billion, roughly 69% of the total | Broadcom residual value support, absorbing declines in asset value | Supported by Broadcom's investment-grade profile |
| Junior (second lien) debt | About $18 billion | Anthropic lease payments | BB plus with positive outlook from KBRA; BB from Morningstar DBRS |
| Blackstone commitment | $9 billion of the junior tranche | Same | Targeted yield above 10% |
Broadcom's role is the part most readers skip. Its residual value guarantee means the senior lenders are not simply betting on Anthropic. They are betting on a chipmaker with an investment-grade rating standing behind the collateral. That structure is why a company with an operating loss could raise $63 billion at all. The senior tranche looks closer to a Broadcom credit than an Anthropic credit. The junior tranche does not, and that is where Blackstone put its money, at a yield that reflects the difference.
KBRA's BB plus with a positive outlook on the junior debt is a small but real improvement over the stable outlook it assigned to a similar arrangement, Project Big Sky, four months earlier. Read that alongside Blackstone's revised corporate-level forecast and the picture is consistent: the agencies see improvement, but not enough to cross into investment grade.
Why a rating is a financing input, not a report card
It is tempting to treat the rating as commentary on Anthropic's business. It is better understood as an input to a fairly mechanical calculation about how the company funds the next several years. The rating is not a judgment about product quality, and it should not be read as a statement that Anthropic will recieve no more funding.
The arithmetic runs like this. Anthropic reported $16.29 billion of revenue in the first half of 2026, up from $1.16 billion a year earlier, alongside an operating loss of $693 million and $8.95 billion of compute and infrastructure expense. The $80.7 billion net loss is mostly an accounting event: about $72.4 billion of it came from non-cash fair value changes on convertible notes, derivative liabilities, and similar instruments. As the company's private valuation climbed, the accounting value of those liabilities rose, and accounting rules pushed that increase through the income statement.
So the operating business is improving while the balance sheet is committing to enormous future payments. Anthropic holds more than $112 billion in cash and cash equivalents, according to the FT, and investors expect it to raise roughly another $100 billion in the IPO. Its S-1 discloses named compute contracts with Amazon, Microsoft, Broadcom, and other providers that add up to roughly $637 billion, and the $800 billion infrastructure plan extends to 2045.
Capital of that size cannot come from equity alone. Equity raises dilute, and public shareholders are less forgiving than private ones about dilution at scale. Debt is the alternative, and the price of that debt is set by the rating. An investment-grade issuer might borrow at a spread that a double-B issuer pays a multiple of. On a balance sheet carrying tens of billions in lease obligations, a couple of hundred basis points of spread is not a rounding error; it is the difference between funding the next tranche comfortably and funding it reluctantly.
It also changes the way counterparties view long-term contracts. A compute agreement with an investment-grade tenant gets more favorable payment terms, longer credit periods, and more room to renegotiate capacity. A speculative-grade tenant faces stricter milestones, more escalation clauses, and more frequent true-ups. None of those terms show up in an API changelog, but they determine how aggressively Anthropic can sign new capacity deals and therefore how much inference and training throughput it can offer.
This is not a report card on the product. It is a pricing signal for the cost of capital. A developer reading the S-1 might assume that a $112 billion cash balance removes all risk. In practice, the cash balance is not enough to cover the forward commitments if the financing markets tighten. The rating matters because it governs how cheaply the remaining gap can be filled.
The accounting details behind the $80.7 billion loss
The $80.7 billion net loss looks terrifying at first glance. Most of it is not an operating loss. It is a non-cash mark-to-market adjustment driven by the rising valuation of the company's private convertible notes and derivatives. When a unicorn raises at a higher valuation, the conversion features on earlier notes become more valuable, and accounting standards require the issuer to record that increase as a fair value loss.
This is one of those cases where generally accepted accounting principles create the opposite of what intuition suggests. As Anthropic became more valuable, its income statement showed a larger loss. The operating business, meanwhile, improved: second-quarter operating income was $974 million, and the first half generated positive operating cash flow of roughly $3.9 billion.
For engineers who work with financial data, the lesson is to separate recurring operating expenses from non-cash valuation charges. The $8.95 billion of compute and infrastructure expense is real spending. The $72.4 billion fair value adjustment is not a cash outflow. That distinction is central to the credit story because rating agencies focus on the company's ability to generate cash and service debt, not on accounting noise caused by the company's own valuation.
The bull case the agencies are weighing
It would be wrong to read the junk forecast as a verdict that Anthropic's business is weak. The case for investment grade rests on numbers that are genuinely unusual.
Revenue grew roughly 1,300% year over year in the first half. Anthropic reported $974 million of operating income in the second quarter of 2026, its first profitable operating quarter in the six periods disclosed in the S-1. It generated positive operating cash flow of roughly $3.9 billion in the first half. Customers spending $100,000 or more on a trailing twelve-month basis grew from over 950 to over 11,900 in a year, and net revenue retention was approximately 900% as of September 30, 2026. A business with that retention profile has a plausible path to covering fixed commitments.
The 900% net revenue retention figure deserves special attention. It means existing customers are spending roughly nine times as much on Anthropic as they did a year earlier. Very few software businesses ever see that metric. For a credit analyst, it reduces the risk that the revenue base will erode quickly if new customer acquisition slows. It also supports the idea that Anthropic could become a large, profitable company within the next few years.
That is the argument bankers have been making to the agencies, and it is not frivolous. But it runs into a structural objection that rating committees tend to take seriously: there is no through-cycle history. Investment grade has historically meant a demonstrated ability to service debt during a downturn. Anthropic was founded in 2021. It has never operated through a demand collapse, a pricing war that compressed margins, or a technology shift that made its assets less valuable.
Moody's made a related point about the broader financing pattern, describing Broadcom as one of the strongest large technology companies while warning that risk would rise if multiple transactions were structured in a short period to meet the sizable funding needs of frontier AI labs. The concern is not any single deal. It is the aggregate. When every major AI lab is trying to lock down datacenter capacity at the same time, the total volume of machine-readable obligations becomes a systemic data point that rating agencies cannot ignore.
How rating agencies think about AI compute contracts
Understanding the rating means understanding how agencies look at compute contracts. These agreements are not like ordinary SaaS contracts. They involve enormous prepayment structures, residual value guarantees, special purpose vehicles, and delivery milestones that can shift by quarters.
Agencies tend to focus on four variables.
The first is debt service coverage, which compares expected cash flow to required principal and interest payments. Anthropic's revenue growth makes this ratio look better today than it did a year ago, but the coverage ratio worsens when analysts add the full amount of committed lease payments into the fixed charge calculation.
The second is asset coverage. If the chips underlying the lease lose value, how much collateral is left after a default? Broadcom's residual value support is designed to answer that question for the senior lenders. The junior lenders are more exposed to Anthropic's own credit, which is why the junior debt receives a projected BB rating while the senior debt is effectively priced off Broadcom.
The third is counterparty quality. A compute contract with Google Cloud is not the same as a compute contract with a startup that offers discounted GPUs. The rating agencies can give credit for strong counterparties, but they also worry about concentration risk. Anthropic's dependence on a small number of hardware and cloud providers is a credit negative even when the providers themselves are investment grade.
The fourth is contract enforceability. Are the payment obligations unconditional? Can Anthropic terminate for convenience? Is capacity guaranteed at a specific date? S-1 filings summarize these agreements, but the actual definitions occupy hundreds of pages. Rating agencies typically assume the worst case if the contracts allow for flexibility.
Structural objections that keep it out of investment grade
The lack of a through-cycle record is only one objection. There are at least four more that rating committees weigh.
First, the assets being financed are technology assets with a useful life that is shorter than the debt maturity being proposed. AI accelerators depreciate quickly, and their resale value depends on generations of hardware that have not been released yet. Broadcom's residual value support helps the senior tranche, but it does not eliminate technological obsolescence. If a new interconnect standard or a software optimization reduces demand for current chips, the collateral value may decline faster than the amortization schedule.
Second, the contracts that back the lease payments are concentrated among a small number of counterparties. Anthropic's compute agreements are with hyperscale cloud providers and specialized GPU clouds. That concentration is operationally important because it gives Anthropic access to capacity, but it is a credit negative because a problem at one counterparty can create a cascading revenue and supply problem.
Third, the structure uses special purpose vehicles to keep leverage off the operating company's consolidated balance sheet to some degree. SPVs are common in project finance, and they are not inherently risky. But they can make the true aggregate leverage harder to observe. Credit investors have been flagging data center and chip financing structured through SPVs for months, so that leverage does not appear cleanly on any single balance sheet. The rating agencies have to estimate how much of a company's future obligations are embedded in these vehicles and how much recourse the parent actually has.
Fourth, there is the efficiency question. Many frontier AI companies have publicly discussed inference-time techniques that reduce the number of chips needed for a given workload. If efficiency gains reduce demand for raw compute faster than the contracts assume, the utilization assumptions behind the lease payments could miss. This is not a problem unique to Anthropic, but it is another reason why an agency might hesitate to assign investment grade to a credit that is effectively a bet on a specific hardware roadmap.
Who is left holding the structure
There is a distributional question buried in the rating. When a deal is investment grade, risk lands on insurers, pension funds, and bond funds operating under conservative mandates. When it is speculative grade, risk lands on high-yield specialists and private credit funds, which demand higher yields and, in a stress scenario, are quicker to sell.
Blackstone's $9 billion commitment in the junior tranche, seeking a yield above 10%, is a clear illustration of how the market prices this. The yield is the compensation for taking Anthropic lease risk without Broadcom's guarantee underneath. If Anthropic's corporate rating comes in at double-B, that junior paper looks appropriately priced. If the corporate rating surprised to the upside, that yield would have been a bargain.
Broadcom carries the other side of the trade. Its residual value support transfers part of the collateral risk off the lenders and onto a chipmaker with a strong rating. The FT's reporting on Project North Star notes that Broadcom's A-range rating was an important factor in getting the financing done at all. That is a reminder that AI infrastructure debt is rarely a pure bet on one company. It is a web of guarantees in which the strongest balance sheet in the room sets the terms for everyone else.
Similar reversals have occured across other AI financings in the past year, and the pattern is nearly always the same: an optimistic private forecast, a preliminary agency discussion, and then a correction toward the middle of the credit spectrum. The correction does not mean the deal fails. It means the people providing the capital price it more honestly.
What to watch between now and pricing
The rating agencies have not spoken. What exists today is preliminary guidance relayed by an anchor investor, and preliminary guidance can move in either direction once audited financials, a price range, and a share count are on the table. A completed IPO that raises close to the reported $100 billion would materially change the liquidity picture the agencies are modeling, and a large equity cushion is exactly the kind of thing that pushes a borderline credit upward.
Three data points will settle the question faster than any commentary.
- The actual rating and outlook from each agency that covers the new debt, including whether any of them assigns a higher rating to the secured tranches than to Anthropic's unsecured obligations.
- The payment schedule behind the compute commitments. A $637 billion contract total says little without knowing how much is due in each year, how much is cancellable, and whether obligations are conditioned on delivery milestones.
- The spread at which Anthropic's first unsecured or lightly secured borrowing prices after listing. That number is the market's own rating, and it will be visible immediately.
The third point is teh one most quantitative developers can start tracking today. Spreads are quoted in basis points, and they move before rating agencies publish anything. If you see a new issue from Anthropic price several hundred basis points above Treasuries, the market has already made its own judgment.
There is a fourth signal that is less obvious: the behavior of the secured versus unsecured spread. If the market prices the secured tranche at a very tight spread and the unsecured corporate debt at a wide spread, it is telling you that the market trusts the collateral and the counterparty guarantees more than it trusts Anthropic's standalone credit. That kind of split can persist for years and is common for project finance platforms.
A practical checklist for engineering leaders
For engineering leaders who consume Anthropic models or build tooling around them, the credit story has practical implications. This checklist is a useful way to prepare.
- Negotiate reserve capacity in writing. If your workloads depend on a specific model family, ask for reserved throughput and committed service levels. A financially constrained vendor may cut unsupported traffic first.
- Build a model abstraction layer. The ability to route between Anthropic, OpenAI, or open-weights models is no longer a nice-to-have. It is a hedge against vendor pricing changes and capacity constraints.
- Monitor unit economics, not just API price. Track cost per successful request, including retries, latency penalties, and cache hit rates. A price cut can be offset by worse reliability or lower inference efficiency.
- Watch the vendor's balance sheet disclosures. After the IPO, Anthropic will report quarterly. Look for changes in committed capex, lease obligations, and cash burn. Those numbers are leading indicators for API pricing.
- Understand the SPV exposure if you use dedicated GPU instances. If you rent bare metal or dedicated capacity from a provider that financed its fleet with project debt, ask whether the financing includes residual value guarantees from a third party.
- Prepare for stronger rate limits. As Anthropic's cost of capital rises, it will prioritize paid, predictable workloads over speculative traffic. Build retry and fallback logic that respects those limits without causing user-facing failures.
This checklist is not about predicting a default. It is about reducing the blast radius if the financing environment changes faster than the product roadmap.
A simple model for the spread drag
For developers who want to quantify the impact of a rating change on capital costs, here is a simple model of the annual spread drag between investment-grade and speculative-grade borrowing on a hypothetical $10 billion note. It is intentionally simplified, but it shows why rating categories matter for anyone who models AI capex.
face_value = 10_000_000_000
years = 10
def annual_cost(spread_bps):
return face_value * (spread_bps / 10000)
ig_spread = 150
hy_spread = 450
ig_cost = annual_cost(ig_spread)
hy_cost = annual_cost(hy_spread)
drag = hy_cost - ig_cost
print(f'Investment grade spread: ${ig_cost:,.0f} per year')
print(f'High yield spread: ${hy_cost:,.0f} per year')
print(f'Annual spread drag: ${drag:,.0f}')
The output is straightforward:
Investment grade spread: $150,000,000 per year
High yield spread: $450,000,000 per year
Annual spread drag: $300,000,000
A single rating notch can create hundreds of millions of dollars in annual interest expense for a company of Anthropic's size. Those costs eventually flow into either the income statement or the prices charged to customers. When you see a company raise prices or reduce free tier limits, this kind of capital cost is often part of the reason.
Common questions from engineering teams
Will a junk rating make Anthropic less safe to use?
No. The rating is about the company's capital structure, not the security of its API. Anthropic has a large cash balance and a fast growing revenue base. A speculative-grade rating increases the cost of future borrowing, but it does not mean the service will disappear.
Does this affect the Anthropic API pricing?
Indirectly. A higher cost of capital puts pressure on the company to improve utilization and pricing discipline. You may see more aggressive tiering, higher prices for low-value traffic, or bigger discounts for committed-use contracts. The effect will show up gradually, not as a single overnight price change.
What should I do if my company depends heavily on one model provider?
Build redundancy. Use an internal gateway that supports multiple providers. Keep evaluation datasets and prompts in a format that is not tied to one vendor. The goal is not to avoid Anthropic, but to make switching possible without rewriting the entire application.
Can the rating improve quickly?
Yes. If the IPO raises an unusually large equity cushion and the company continues to grow at triple-digit rates, the agencies could revise the outlook to positive and upgrade within 12 to 24 months. The Blackstone forecast is a snapshot, not a life sentence.
How do I track this going forward?
Follow the rating agency reports, the company's quarterly earnings, and the new issue calendar. The spread on the first corporate bond issue is the single most transparent market signal available.
The signal inside the forecast
A junk rating is not a judgment that Anthropic is failing. The company grew revenue by more than a factor of fourteen year over year, turned an operating profit in the second quarter, and sits on a cash pile larger than the entire revenue base of most public software companies. Plenty of excellent businesses have spent years in high yield while they built durable cash flow.
What the Blackstone reversal signals is narrower and more useful. The credit market is separating the question of whether Anthropic will be a huge business from the question of whether its obligations should be held by conservative institutions. The answer to the first is increasingly yes. The answer to the second, at least for now, is: not yet, and not at investment-grade prices.
For anyone watching the AI buildout, that distinction is the real story coming out of the IPO. Anthropic's valuation will be set by people buying a growth story. Its borrowing costs will be set by people asking who pays if the growth arrives late. The gap between those two groups, and how it narrows over the next few quarters, will shape how much computing capacity the company can actually commit to. The FT's report is the first clear reading of how wide that gap currently is. For engineers and infrastructure buyers, the practical takeaway is equally clear: monitor the ratings, track the capital costs, and design your AI stack to survive a world where one large vendor has to borrow at junk prices.
Top comments (0)